Transcription
All right, so this week in AI is going to be a very interactive one. A lot of things you can either try yourself right away for free, or things that are really worth looking at that we haven't seen before, like a Vibe coding game jam on Twitter and more. And this week's episode of AI news you can use shows that pulls all together all the AI madness into one video that filters out all the releases that are just distant promises or inferior versions of what we have already. So, if the little intro is being done here, let's get to the first piece of AI news that you can use. Okay, so for our first segment, we'll be looking at Gemini 2.0 personalization, their experimental model that uses your Google search history as a context for the chatbot.
Look, if you've been following this show, usually there's one story that I use right in the beginning; we also use it for the packaging of the video, the thumbnail, and the title that you see on the homepage. And most of the times, that's a story that excites me personally the most. And in this week's video, it is this one, because I don't know about you, but for me personally, when I had a lot of conversations around the potential of AI and all of these tools, there's usually this one concept that really resonates with people, and people just get it; they're also afraid of it, but it's the concept of it having access to your data, to your search history, to your browser history. And this is the first version of a product on the market at all that I see doing this—a pretty fearless thing to be honest, not very typical for Google. We'll talk about that, but now let's have a look at the model, which is accessible through gemini.com. You can access this on a free plan, as you can see right here; I'm on a free plan. You could try this out yourself, but you do have to be logged in with a Google account; that's where it takes the Google search history from.
But basically, this allows you to tab over to this personalization model, which uses your search history in Google that is tied to this account to inform the model. So it's like setting up custom instructions, but it happens automatically, and it happens from all your searches. Now, if you've been following the channel, you know I'm not the biggest fan of these automated context tools, like memories in ChatGPT, just because you're not in control. For casual users, it's a great feature that can greatly improve the experience, but if you look at my ChatGPT history, you might just think it's multiple people sharing the account. Just one moment; I'm using it to analyze thought patterns, and the next moment I'm using it to strategically plan for my business. If it uses the context of one for the second, it's not a good match. And in this case, it's the way they presented it in their blog post, and the way I understand it is pretty much them looking at this, but the extended version: your search history. So as you can see, there are some AI tools here; I was looking at details on market cap of BU, checking some flights, and this is a quest inside of a computer game that I've been playing recently, Kingdom Come Deliverance. To it's it's incredible; can only recommend, by the way. But as you can see, this is all all super disconnected, so I wonder how this works. And let's ask it the most essential thing, which is: "Tell me about myself, be detailed." And now let's see what it knows about us. There's no custom instructions, no other context set up here; it's just pulling in my history. Okay, I've showed an interest in AI; I've been tracking investments—fair enough; software dev and online tools, of course. Wow, it really has a lot here. Oh my God, I can pick up that I'm into video games. Wow, this is more powerful than I would imagine. Wow, I got to say this digs deep. I guess the search history does say a lot about yourself.
Now, now after reading this in detail, I got to say there's actually a point in here that I also don't want to share in the video. It was a French search term relating to a weird topic on AI that I was researching, and I absolutely would not want that in my context. But for this use case—you're telling it about yourself—it is super interesting. And the first thing that comes to mind here is that this would be super useful to improve your personal context. We've spoken about this before, but I'm a big advocate of you having a document that basically describes yourself, so if you're prompting AI, and if you want to tell it what you're about, you can just copy-paste that document. This seems really helpful in crafting that, but I also have to say it's a scary feature that you have access to. Now, if you're in the US—as you might know, I live in Portugal, Europe—here I didn't have access to this, so I needed to use a VPN to get this. It's available both on the free and the paid account, so if you're in the US or you have a VPN, you can just try a prompt like this and see if the description of Google about yourself is accurate. In my case, I have to say I'm surprised it actually is; I thought that the searches were really scattered, but this is good stuff. And I want to make one final point before we move move on, and that is: two years ago, Google was super afraid to ship anything AI related that might ruffle some feathers. Remember the founding fathers that were ethnically diverse? Well, this is kind of on the opposite end of the spectrum; it's using your search history, and if it went even a step further, it would use all of your email conversations and maybe even your personal messages. This is a bit of a slippery slope, and this channel is here to show off the use cases and not to have this ethical discussion around it. I just want to point out that between this and the image editing tool they released last week—by the way, separate video on various use cases of that should be up on the channel soon—allows it to upload an image that has watermarks, and with free words, remove the watermarks; it will do that perfectly, which is a very different approach from censoring everything and trying to make every generation as unoffensive as possible. Just an interesting development that I kind of wanted to highlight here. And now let's move on to the next story, which is going to be brief, but it also relates to this Gemini platform here, and that is then basically catching up with some of the competing products; they're including a canvas feature now, which also includes code preview, similar to the way ChatGPT has it. Claude was first on that with their artifacts, by the way. One difference between Gemini and some of the competition is that you have to enable this canvas here at the bottom, and as you can see for 2.0 pro experimental, it's grayed out; matter of fact, it's grayed out for all models except the flash right now, which, as you can see now, has a canvas feature like most other LLMs now, too. There you go; can run Python code. Beautiful about the Deep research feature now; it uses their 2.0 models; it's better than the Deep research before. They, by the way, changed interface just like the way ChatGPT and Grok and Perplexity have it now, too. But from our brief initial tests on this, it's still inferior to ChatGPT and sort of on the level of Grok and Perplexity now, whereas before they were behind. So really, all of these updates are them just catching up, which is good to see, cuz all of these features on the other platforms are powerful, and this update mostly catches them up. And for anybody skeptical of this invading your privacy, well, if you weren't aware, Google search is already using your search history to improve your Google search suggestions, so if you're using Google, this is not something you should really be worried about. I think the discussion should be more about the direction this is going in rather than this particular implementation, which is certainly a step of using more and more of your data to customize these experiences.
And then there's one new feature which isn't available in any other LLM platform, and that is an audio review. When you run one of their new deep researches, audio overview is something they have inside of their other app called NotebookLM, and it basically turns a very long document into a podcast-like conversation. It sounds incredible; matter of fact, Google's implementation of this has been the best in the market for a while now, and now they're bringing it to Gemini advance. So if you run deep researches, you can turn the big research document into a concise audio podcast with the click of a button. Now that the Deep research is done, looking good; I like that it separates out the sources it got but didn't use. We can generate an audio overview like so. Now there we go; after about 3 minutes, we have a 5-minute overview of this entire research. Let's have a listen. Okay, so today we're taking a deep dive into finding the cheapest flight from Tokyo to Lisbon, yeah, right around April 20th, exactly April 20th, and you've already done a bunch of research on those flight comparison websites. I mean, it sounds great. Honestly, in this case with the flights, it's probably better to just kind of look at the output and look for the part that you care about, but I can see this being super handy, and no other LLM offers a feature like this right now. You kind of have to take the output and throw it into NotebookLM to do this. Really nice to have it directly in here. All right, so as you can see, Google has been shipping recently, and I'm here for it; they're really stepping up their AI game; might be even worth a dedicated video soon.
Okay, and as we touched on NotebookLM, I also want to add that there's a new feature in there which I've seen in other apps, but I think it's really nice that it's included in NotebookLM just because it makes so much sense. Usually the way you use NotebookLM is you drop like dozens of sources in there, and then you can chat with them, generate podcasts from them, and now they added another thing you can do: you can also visualize them in a mind map. So as you can see from our little test run right here, you work with it as expected, but now, additionally, you get a mind map like this, where just with a click of a button you can keep expanding this, and you can see how different ideas relate to each other, which can be super useful if you add a lot of data into this. And it's something that's not available in any one of the major LLM platforms like GPT, Gemini, or Claude right now. And that's really it. If you're a NotebookLM user, you can now enjoy these mind maps. On to the next one, which is another Chinese model; these don't stop. Now, here's the thing: if you want to try it out, well, the website doesn't offer translations. Sure, you could use Google Translate and get something that looks like this for yourself, but then quickly, if you want to try this out, you're going to encounter another problem: you really need a Chinese phone number to try this out. Now, is this new model even worth your attention? Well, it's a release by the Chinese tech giant Baidu, and concretely, they released two models: Ernie 4.5 and Ernie X1. Ernie 4.5 is really a GPT-4, maybe sort of a GPT-4.5 competitor, but not really from what I've seen so far. And X1 is a DeepSeek or AlphaMini competitor, aka a thinking model. So while these releases are capable, benchmarks look good, sure, the main selling point here is the price through the API. If you look at this graph right here, you will see that it scores on par with GPT-4.5. They also claim it's a model that specializes in empathy, but if you compare compared to GPT-4.5, well, the input tokens are less than 1% of the price when you compare it to GPT-4 at 55 cents per 1 million input tokens and 220 for 1 million output tokens; it's roughly 1/5th of the price. But I have to be honest, I really wanted to try this, but I just don't have a Chinese phone number. Maybe I have to do something about that to test some of these models that keep coming out. From the outputs I have seen, I don't know, better than GPT-4? Sure. Better than 4.5? I don't think so. But the essence of this is basically new Chinese models that are super capable, and the X1 model is even half the price of R1, which was already a fraction of the price of all the Western models. So this trend just continues, and I guess I should get a Chinese phone number. But there you go; that's all the info I have right now. On to the next one.
Okay, so this is a story that has been going on on Twitter for the past few weeks. I'll give you a 15-second summary, and then we'll get into talking about this Vibe coding game jam that is happening. But basically, Levels.io, here, online entrepreneur, builder, naturally curious badass, Christopher Columbus of online businesses—whatever you want to call him—this guy basically tries new technological tools all the time, turns them into businesses. He Vibe coded this little game with Cursor, which, if you're not familiar with Vibe coding, it's basically using AI assistance to code for you with a tool like Cursor and a model like Sonnet 3.7 in the background. And it created this entire little flight simulator game, which, by building it in public to his Twitter audience, he managed to get over 320,000 people to fly inside of the game. He started selling ad spots and made over $887,000 in a month, which basically would add up to $1 million AR in a year. This one inspired a lot of people and also pissed off a lot of people, especially full-time game devs. But I personally think this is absolutely amazing and a great show of creativity. And now he's organizing the 2025 Vibe coding game jam, which, amongst others, has Andrej Karpathy as one of the judges, and I just wanted to point it out here: if you ever wanted to create a game with AI, this is a great opportunity to kind of pick that up. I think from the upload of this video you might have another four days, or at the very least you can follow the story and see what people create with AI over the coming week. I personally think this is fascinating. I also understand the people who pour years of their life into building a game, and then they don't get any distribution, and then Peter just throws something up over the weekend, and hundreds of thousands of people use it, and he makes almost $100K of it in no time. But that's the world we live in; it's not just about the idea, it's also about selling it. So yeah, Vibe coding progressing every week; love to see it. On to the next story, which is our prompt of the week in the newsletter actually, but I also wanted to feature it here because if you're getting into Vibe coding, this prompt right here is actually a lifesaver. And I really also wanted to feature it to the YouTube audience because I know that differs slightly with the newsletter audience. By the way, if you're not on our newsletter, it's free and you should absolutely be there; link in the description below. You get some of my favorite prompts and apps as you sign up, and then you get one email a week that includes a few segments; one of them is our prompt of the week, which this week is this one from Ted Verbal. It's basically a fantastic way to troubleshoot things inside of something like Cursor or Windsurf when you're stuck. So there you go; that's a little Vibe coding segment for this week. I've been making things, too, though. This is my little self-developed ChatGPT prompt presets templates creator. Look at that; I made this thing in about an hour and still use it. You can add things here, and in this way I can manage my most used prompts within this little extension that I just Vibe coded myself. Also, if you want to learn more about that, you might know we have an AI learning community, and in there I held a 90-minute lecture on how to get started with Vibe coding yourself for non-technical people, concretely, claw coder. While that has been a really popular lecture inside of our paid community, and it's the type of thing we do there on a weekly basis.
Okay, next up, we had Nvidia GTC happening this week; this is their big developer conference, and during the keynotes they present all their new partnerships and updates and how things are going and what they're about to release. They open-sourced some frameworks; they announced a lot of big partnerships, like partnering with GM to enable full self-driving. But I think from the entire keynote—which you can find a fantastic fantastic summary of here on the Cenut channel; I'll link it below; it's just 15 minutes and kind of wraps up everything—if you enjoyed this video, you would probably enjoy this one, too. But the one thing that really caught so many people's attention across the team and also other people I've been talking to is this collaboration with Disney Research and Google DeepMind. Here in the end, they basically built a Star Wars robot that they showed off on stage. Just have a look at this guy; isn't this incredible? I mean, okay, okay, what's the use case? Can you buy it? Like, I don't know. And no, you can't buy it, but I don't know; I just feel like sometimes it's good to stop and appreciate how far technology has come, and and you know, people have been trying to build these little things forever, fans of Star Wars, etc. But the details on this thing and the motion—obviously this uses a lot of the hard and software they've been presenting in the rest of the presentation; for example, it has a brand new physics engine that allows it to, well, walk across the stage like a little duck. And Nvidia even open-sourced the software that runs their Groot humanoid robots. But I just kind of wanted to highlight this little guy here, cuz I don't know; I've been following these stories all my life in tech, and it still amazes me, and this is just one of those moments, and I wanted to share it. Okay, on to the next one.
Okay, so the next one is a website for something that has been around for a while, and that's basically people testing LLMs on a benchmark that lets the AI build different structures in Minecraft. Minecraft, you might be familiar; it's the world's most played game, I believe. Ah, okay, it's only second because the first place is held by Tetris. Fair enough. But basically, challenges these models to build structures inside of Minecraft, and in this site you can rate them yourself, and then you can look at a leaderboard of the models who are best at this, and it's really fun. So let me show you which one of these looks more like a table to you. Well, I would say this one, so I vote A. Next comparison: amio strip. The heck is that? Aha, okay, it's the logo of every 10th tech company these days, I see. Well, definitely this one made by GPT-4.5 versus Gemini over here. Next comparison: person meditative zengarden with a pagoda. Oh, definitely B; Gemini 2.0 flash wins. I think you get the point here. If you want, you can tab over to the leaderboard to see which model performs best at this. Oh, it's Claude 3.7 Sonnet. Who would have thought? Who would have thought? Not me. Well, me; I would have thought; already in the release video made it quite clear that this is without dispute the best coding model out there. So if you follow the channel or Gen, you probably knew this. Nevertheless, I instantly bookmarked this page, cuz there are so many benchmarks, but honestly, this is one of my favorites cuz it's so visual, and you can actually see the model doing something and planning its actions rather than some arbitrary number with test cases in the background that you don't even get to review, and most of the time you're not sure if the model was trained on exactly those questions. So there you go; get a few votes in, bookmark this if you want, and now let's look at the next piece of AI news that you can use.
Okay, so the next release is something that you probably missed because it's an AI that generates 3D models that is being integrated into what's essentially a kids' game. But this is the first time that, at least I'm seeing some of these generative 3D models being integrated into something that is going to not hit just millions but hundreds of millions of people, because what I'm talking about here is the game Roblox, and their generative AI system is called Cube. It can basically create 3D models from text prompts or voice prompts, like we've seen before on the show. But by the end of this week, it's actually going to be integrated into their game. Now, if you're not familiar with Roblox, a brief overview of some numbers: the game has 85 million daily active users right now; that's a 380 million monthly active users across the world. Now, the reason I said you probably haven't heard of this news or the game is because almost half of those users are under the age of 13. But here's the interesting part: out of those 380 million monthly active users, 2.5 million are developers, because inside of this game you can create little mini-games for yourself; matter of fact, there are over 40 million games and experiences in there. So basically, around half a percent of the players don't just enjoy the game; they also build experiences in there, and tools like this are going to open up so many creative freedoms. Well, you don't need to know 3D modeling, too. 3D modeling while creating your own little gamer experience. And while that might not be the biggest impact today, this is really a thing that is just going to be taken for granted by today's kids and teenagers. Yeah, speaking 3D models and games into existence—no big deal for them; it's always been that way. And I think the long-stream effects of that are actually incredible because, as opposed to things like social media where you're more on the consumer side, here you're actually on the developer, the creator side. And if you want a pink buggy, well, you just say you want one, and it appears. Pretty amazing use of technology for today's youth, in my opinion. So I wanted to show you, and that's really all we've got for this week: a bunch of things that you can try. Go and check out what Gemini thinks about your search history, or go and rate some Minecraft models. And I'll be back next week with another episode of AI news you can use. And I myself, I'm heading off to Dublin in about 5 minutes here, about to hold a little corporate workshop on how hyper-personalization through the use of all of these tools that we keep looking at week by week is going to affect the company that I'll be working with. I'm super excited; teaching is always what I enjoyed the most, and as time passes I get to do more more of that. So thank you so much to even you for all the support on this YouTube channel, and yeah, that's pretty much all I got for this week. See you soon, bye.